Meteorological downscaling method based on space-time fusion and physical constraint

Abstract

The invention provides a meteorological downscaling method based on space-time fusion and physical constraint, and belongs to the technical field of meteorological downscaling, and the method comprises the steps: carrying out the preprocessing of multi-source meteorological related data and a high-resolution meteorological truth value, and constructing a training data set; an improved U-Net model is constructed, spatiotemporal features and multi-source auxiliary features are obtained through a multi-branch feature extraction unit, high-resolution information is recovered through fusion and decoding, and an attention enhancement module is embedded to highlight a key area; a model is trained through a training data set, parameters are optimized by adopting a loss function fusing topographic features and physical rules, and prediction error differentiation constraint on a complex area and violating the physical rules is achieved; and preprocessing target low-resolution data, inputting the preprocessed target low-resolution data into the model, and outputting high-resolution meteorological data and a physical attribution result. The problems that in the prior art, multi-source meteorological data fusion is insufficient, downscaling precision of a complex terrain area is insufficient, and prediction errors violating physical laws are lack of effective constraints are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of meteorological downscaling, in particular to a meteorological downscaling method based on spatio-temporal fusion and physical constraints. BACKGROUND

[0002] In the field of meteorological downscaling, with the development of satellite, radar monitoring and numerical prediction technologies, meteorological observation and prediction data with certain spatio-temporal resolution can be obtained at present. Through the quality control technology of integrating artificial intelligence and mathematical statistics methods, the processing capacity of observation data has also been improved. In terms of model construction, there are various models for meteorological analysis, such as numerical prediction models that form high-precision grid analysis fields through processing and data assimilation of massive multi-source observation data, and then predict future weather evolution.

[0003] Meteorological research and application have increasingly high requirements for the precision and resolution of meteorological data. In order to meet the needs of fields such as urban disaster prevention and reduction, fine meteorological service for agriculture, and hydrological and meteorological prediction of river basins, downscaling technology has become a research hotspot. At the same time, fusing multi-source data (such as satellite cloud images, radar echoes, terrain data and conventional meteorological data) and combining physical mechanisms and artificial intelligence technology to construct more accurate meteorological models are important directions for future development. Similar to AI meteorological application, implementing an artificial intelligence weather prediction large model demonstration plan and developing a short-to-medium-term prediction large model that fuses atmospheric physical mechanism constraints are measures to comply with this trend.

[0004] However, the existing technology still has deficiencies in meteorological downscaling. On the one hand, traditional downscaling methods cannot fully utilize the complementary information of multi-source data and cannot accurately depict the influence of complex terrain and other factors on meteorological elements. For example, in complex terrain areas, due to the lack of in-depth mining of the relationship between terrain features and meteorological elements, the description of local meteorological conditions after improving the resolution of meteorological data deviates greatly from the actual situation. On the other hand, the existing models lack effective constraint mechanisms when dealing with prediction errors that violate physical laws, which reduces the reliability of the prediction results in some extreme or special meteorological scenarios. SUMMARY

[0005] In order to overcome the deficiencies of the existing technology, the purpose of the present application is to provide a meteorological downscaling method based on spatio-temporal fusion and physical constraints, which realizes accurate downscaling of low-resolution meteorological data by multi-branch feature extraction, spatio-temporal fusion, attention reinforcement and loss function optimization that fuses terrain and physical rules, improves the prediction accuracy and physical rationality of high-resolution meteorological data in complex terrain areas and special meteorological scenarios, and can output physical attribution results to clarify the influencing factors.

[0006] To achieve the above purpose, the present application provides the following scheme:

[0007] A weather downscaling method based on spatio-temporal fusion and physical constraints, comprising the following steps:

[0008] S1, preprocessing multi-source weather related data and high resolution weather true value, constructing training data set; the preprocessing includes standardization and feature extraction;

[0009] S2, constructing improved U-Net model, the model obtains spatio-temporal features and multi-source auxiliary features through multi-branch feature extraction unit, restores high resolution information through fusion and decoding, and embeds attention reinforcement module to highlight key regional features;

[0010] S3, training the model through the training data set, adopting loss function of fusion terrain feature and physical rule to optimize parameters, realizing differentiated constraint on prediction error of complex area and violation of physical law;

[0011] S4, after the target low resolution data is preprocessed in step S1, the trained model is inputted, and high resolution weather data and physical attribution results are outputted.

[0012] Preferably, in step S1, the multi-source weather related data includes low resolution weather data, terrain data and multi-source observation data.

[0013] Preferably, the low resolution weather data includes one or more of wind speed, temperature and irradiance, the terrain data is digital elevation model data, and the multi-source observation data includes satellite cloud image and radar echo.

[0014] Preferably, in step S1, the standardization is adaptive normalization, specifically:

[0015]

[0016] Wherein, x' is the standardized data, x is the original data, μ(x W ) is the mean value in the sliding time window W, W is 24 hours, σ(x W ) is the standard deviation in the sliding time window W;

[0017] The feature extraction includes calculation of terrain roughness, and the formula is:

[0018]

[0019] Wherein, R is the terrain roughness, β is the slope, α is the slope direction, is the gradient operator.

[0020] Preferably, in step S2, the improved U-Net model comprises a multi-branch feature extraction unit, a space-time fusion module, an attention reinforcement module and a scene adaptive decoding unit; wherein the multi-branch feature extraction unit comprises three parallel branches:

[0021] The meteorological space-time branch extracts spatial features through 3×3 convolution and captures time sequence dynamics by combining the Transformer self-attention module. The time sequence attention weight calculation formula is:

[0022]

[0023] Wherein, α t is the time sequence attention weight, Softmax is the normalization function, Q t is the time sequence query vector, K t is the time sequence key vector, T is the matrix transpose symbol, d k is the vector dimension;

[0024] The terrain branch uses a residual block to process terrain derived features and enhance the feature extraction capability of complex terrain areas.

[0025] The multi-source observation branch extracts features from satellite cloud images using 7×7 convolution and radar echoes using 5×5 convolution to adapt to the distribution characteristics of different observation data.

[0026] Preferably, the space-time fusion module is used to fuse the features output by the multi-branch feature extraction unit, and the spatial and time sequence features are associated learned through a cross-attention matrix, with the formula being:

[0027]

[0028] Wherein, M st is the space-time cross-attention weight matrix, F s is the spatial feature vector, T is the matrix transpose symbol, F t is the current time sequence feature vector, σ st is the space-time feature covariance, which is adaptively learned from training data; n is the number of historical time points, which is 6; F t,i is the i-th historical time sequence feature vector.

[0029] Preferably, the attention reinforcement module is used to highlight key area features, and the feature importance of different areas is adjusted by a spatial attention weight, with the formula being:

[0030]

[0031] Wherein, W sFor spatial attention weight, Sigmoid is the activation function, Conv is 3*3 convolution operation, Cat is feature splicing operation, MaxPool is maximum pooling operation, AvgPool is average pooling operation, F is input feature map, is the terrain roughness gradient.

[0032] Preferably, the scene adaptive decoding unit is used to dynamically adjust the decoding strategy according to the weather scene, and the branch activation function is determined by scene classification, and the formula is:

[0033]

[0034] Wherein, P c is the activation probability of the cth scene, exp is the exponential function, MLP is the multi-layer perception mechanism, F stat is the statistical feature vector of the current data, F stat,i is the reference statistical feature of the ith scene.

[0035] Preferably, in step S3, the loss function calculation formula of the fused terrain feature and physical rule is:

[0036]

[0037] Wherein, L is the comprehensive loss value, y is the high-resolution weather true value, is the predicted value, R is the terrain roughness, tanh is the hyperbolic tangent function, y wind is the predicted wind speed, y wind,ref is the reference wind speed, y rad is the predicted irradiance, y rad,max is the theoretical maximum irradiance, which is calculated by the solar elevation angle.

[0038] Preferably, in step S4, the terrain contribution degree of the physical attribution result is calculated by the formula:

[0039]

[0040] Wherein, C R is the terrain contribution degree, the value range is 0-1; W s is the spatial attention weight, is the terrain roughness gradient, y is the high-resolution weather true value, is the predicted value, is the absolute error of the predicted value and the true value.

[0041] According to the specific embodiments provided by the application, the following technical effects are disclosed:

[0042] (1) This invention constructs an improved U-Net model containing multi-branch feature extraction units, extracts meteorological spatiotemporal features, terrain features and multi-source observation features respectively and fuses them, realizing the deep integration of multi-source meteorological related data, solving the problem that traditional methods cannot make full use of the complementary information of multi-source data, and improving the data utilization rate.

[0043] (2) This invention highlights key regional features by embedding an attention enhancement module and incorporates terrain features into the loss function for weighted optimization, thereby enhancing the ability to capture features of complex terrain regions. This makes the prediction results of high-resolution meteorological data in complex terrain regions more realistic and improves the downscaling accuracy of complex terrain regions.

[0044] (3) This invention constrains prediction errors that violate physical laws by using a loss function that integrates physical rules, and dynamically adjusts the decoding strategy by combining a scene adaptive decoding mechanism, thereby ensuring the physical rationality of the prediction results and improving the reliability of the prediction results under special meteorological scenarios. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of a meteorological downscaling method based on spatiotemporal fusion and physical constraints according to the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] Example 1

[0050] like Figure 1 As shown, this invention provides a meteorological downscaling method based on spatiotemporal fusion and physical constraints, comprising the following steps:

[0051] S1, preprocessing multi-source meteorological related data and high-resolution meteorological true value to construct a training data set; the preprocessing includes standardization and feature extraction;

[0052] S2, constructing an improved U-Net model, the model obtains spatio-temporal features and multi-source auxiliary features through a multi-branch feature extraction unit, restores high-resolution information through fusion and decoding, and embeds an attention reinforcement module to highlight key regional features;

[0053] S3, training the model through the training data set, using a loss function that combines terrain features and physical rules to optimize parameters, and realizing differentiated constraint on prediction errors in complex areas and areas that violate physical laws;

[0054] S4, inputting the target low-resolution data after preprocessing in step S1 into the trained model to output high-resolution meteorological data and physical attribution results.

[0055] Specifically, in step S1, the multi-source meteorological related data includes low-resolution meteorological data, terrain data and multi-source observation data. Continuous data covering complex terrain areas (such as mountainous areas, valleys and urban building groups) in the past 5 years are selected to ensure that the training set contains diverse terrain and meteorological scenes, such as sunny days, precipitation, high temperatures and cold waves.

[0056] Further, the low-resolution meteorological data includes at least one of wind speed, temperature and irradiance, with a resolution of 10km x 10km and a time interval of 1h, and is derived from the output of a regional numerical weather prediction model. The terrain data is digital elevation model (DEM) data with a resolution of 30m x 30m, which is used to extract terrain features. The multi-source observation data includes satellite cloud images and radar echoes, among which the satellite cloud images have a resolution of 1km x 1km and are updated every 30 minutes; the radar echoes have a resolution of 0.5km x 0.5km and are updated every 6 minutes, which are used to supplement fine observation information. In addition, the high-resolution meteorological true value is obtained by using ground-based intensive automatic station observation data after quality control, i.e. removing outliers and filling missing values, as the supervision label for model training.

[0057] In addition, in step S1, the standardization is adaptive sliding window normalization, specifically:

[0058]

[0059] where x' is the normalized data, x is the original data, μ(x W ) is the mean value in the sliding time window W, W is 24 hours, σ(x W ) is the standard deviation in the sliding time window W;

[0060] The feature extraction includes calculating the terrain roughness, the formula is:

[0061]

[0062] wherein R is the terrain roughness, β is the slope, the value range is 0°-90°, α is the aspect, the value range is 0°-360°, is the gradient operator, a 3x3 Sobel convolution kernel is used for calculation, the horizontal direction kernel is:

[0063]

[0064] the vertical direction kernel is its transpose, in order to enhance the spatial continuity of the terrain features. Wherein, the value range of R is 0-1, the greater the value, the more complex the terrain, to provide the basis for subsequent attention reinforcement and loss weighting.

[0065] In step S2, the improved U-Net model includes a multi-branch feature extraction unit, a space-time fusion module, an attention reinforcement module and a scene adaptive decoding unit; wherein the multi-branch feature extraction unit includes three parallel branches:

[0066] The meteorological space-time branch extracts spatial features through 3x3 convolution, extracts the spatial distribution features of low-resolution meteorological data, and combines the Transformer self-attention module to capture the time sequence dynamics, the time sequence attention weight calculation formula is:

[0067]

[0068] wherein α t is the time sequence attention weight, Softmax is the normalization function, Q t is the time sequence query vector, K t is the time sequence key vector, T is the matrix transpose symbol, d k is the vector dimension; the input time sequence length is 6, that is, the current time characteristics are modeled by using the previous 6 hours of data, and the influence weight of the historical time is dynamically allocated through α t , such as the data weight of the first hour before the sudden rainfall is higher.

[0069] The terrain branch uses residual blocks to process terrain derived features, and enhances the feature extraction capability of complex terrain regions; specifically, 2 layers of residual blocks are used to process the terrain roughness R and the slope β, each residual block contains 2 3x3 convolution layers, and ReLU activation is added in the middle, in order to solve the gradient vanishing problem of complex terrain region feature extraction.

[0070] The multi-source observation branch extracts features by using 7*7 convolution for satellite cloud images and 5*5 convolution for radar echoes, and adapts to the distribution characteristics of different observation data. Specifically, 7*7 convolution is used to extract the distribution characteristics of large-scale cloud systems for satellite cloud images; 5*5 convolution is used to extract local heavy rain characteristics for radar echoes, and the output features are unified in dimension by 1*1 convolution and then spliced with other branch features.

[0071] The spatio-temporal fusion module is used to fuse the features output by the multi-branch feature extraction unit, and the cross-attention matrix is used to realize the correlation learning of spatial and temporal features, and the formula is:

[0072]

[0073] Wherein, M st is a spatio-temporal cross-attention weight matrix, specifically a 7*7 weight matrix, which is composed of 6 historical time points and 1 current time point, and the dynamic fusion of spatio-temporal features is realized by matrix penalty. For example, when the wind speed in mountainous area is affected by terrain and monsoon, M st will automatically increase the weight of the corresponding features of the two. F s is a spatial feature vector (convolution output from the meteorological spatio-temporal branch), T is a matrix transpose symbol, F t is a current time sequence feature vector, σ st is a spatio-temporal feature covariance, the initial value is 0.1, which is updated by training data self-adaptive learning, and is used to balance the scale difference of spatio-temporal features; n is the number of historical time points, which is 6; F t,i is the i-th historical time sequence feature vector.

[0074] The attention reinforcement module is used to highlight the key area features, and the formula is:

[0075]

[0076] Wherein, W s is a spatial attention weight, Sigmoid is an activation function, Conv is a 3*3 convolution operation, Cat is a feature splicing operation, MaxPool is a maximum pooling operation, AvgPool is an average pooling operation, F is an input feature map, is a terrain roughness gradient.

[0077] The scene adaptive decoding unit is used to dynamically adjust the decoding strategy according to the meteorological scene, and the branch activation function is determined by scene classification, and the formula is:

[0078]

[0079] Wherein, P cThe activation probability of the cth scenario is exp, the MLP is a multi-layer perception mechanism, F stat is a statistical feature vector of the current data, F stat,i is a reference statistical feature of the ith scenario.

[0080] In step S3, the loss function formula for fusing the terrain features and the physical rules is:

[0081]

[0082] where L is a comprehensive loss value, y is a high-resolution meteorological true value, is a predicted value, R is a terrain roughness, tanh is a hyperbolic tangent function, y wind is a predicted wind speed, y wind,ref is a reference wind speed, specifically a 95% quantile value of the training data, i.e., an extreme wind speed reference value, y rad is a predicted irradiance, y rad,max is a theoretical maximum irradiance, which is calculated from the solar elevation angle.

[0083] Specifically, the physical rule constraints include a wind speed constraint and an irradiance constraint, where the wind speed constraint is: when the predicted wind speed y wind is less than the reference wind speed y wind,ref , max(0, -y wind / y wind,ref ) is positive, and the exp term amplifies the error;

[0084] The irradiance constraint is: when the predicted irradiance y rad is greater than the theoretical maximum irradiance y rad,max , max(0, (y rad -y rad,max ) / y rda,max ) is positive, and the exp term penalizes the predicted value that exceeds the physical limit.

[0085] The training strategy includes:

[0086] Optimizer: Adam optimizer is used, the initial learning rate is 0.001, and it is attenuated by 10% every 5 rounds;

[0087] Iteration stopping condition: when the mean absolute error (MAE) of the validation set does not decrease for 3 consecutive rounds, and the fitting degree (correlation coefficient R 2 ) of the high-resolution true value is greater than or equal to 0.85, stop;

[0088] Regularization: L2 regularization term (weight 0.0001) is added to avoid overfitting, and the training batch size is set to 32 (balance memory occupation and convergence speed).

[0089] In step S4, the target low-resolution meteorological data (10kmx10km) is spliced with the corresponding terrain data and observation data after standardization and feature extraction in step S1. The model outputs: 1kmx1km high-resolution meteorological data (wind speed, temperature, irradiance), the spatial grid is aligned with the original low-resolution grid through bilinear interpolation, and the time interval is kept at 1 hour. The formula for calculating the terrain contribution degree of the physical attribution result is:

[0090]

[0091] wherein C R is the terrain contribution degree, the value range is 0-1, and the larger the value indicates that the terrain is the main source of error; W s is the spatial attention weight, is the terrain roughness gradient, y is the high-resolution meteorological true value, is the predicted value, is the absolute error of the predicted value and the true value.

[0092] In order to verify the downscaling effect of the application under different meteorological elements (wind speed, light intensity) and typical landforms (plain-near sea), a near-sea wind power plant in Hebei Province and a photovoltaic power plant in Shandong Province were selected as the experimental target area.

[0093] Example 2

[0094] The selection of the experimental target area focuses on the representativeness of typical landforms and meteorological element characteristics: the terrain of a near-sea wind power plant in Hebei Province is flat and the elevation difference is less than 20m, the annual average wind speed is 6-7m / s, and it is significantly affected by the sea-land wind circulation. The selection of this area aims to verify the restoration ability of the application to the fine structure of the wind speed field (such as the sea-land wind coupling boundary layer wind speed gradient); the annual average total solar radiation of a photovoltaic power plant in Shandong Province reaches 5400MJ / m 2, the diurnal and seasonal variation of cloud cover is intense, selecting this area can focus on verifying the robustness of the application in the scene of significant light field gradient and cloudy disturbance. In terms of data sources, the input field uses ECMWF ERA5 reanalysis data, with a spatial resolution of 0.25° (about 28km x 28km) and a time resolution of 1 hour, containing wind speed (10m height) and total radiation (surface) elements; The true value field uses ERA5-Land high-resolution reanalysis data, with a spatial resolution of 0.1° (about 11km x 11km) and a time resolution of 1 hour, which has an error of less than 5% verified by ground observation stations, and can be used as the true value of the downscaling target. Data division, the training set covers 8784 times from January 1, 2024 to December 31, 2024, covering a complete annual cycle to include seasonal weather characteristics; The independent verification set is 2184 times from January 1, 2025 to March 31, 2025, covering the wind speed fluctuation and radiation enhancement period from the end of winter to the beginning of spring, ensuring that the model generalization evaluation has no data leakage. In the preprocessing stage, the input field and the true value field perform step S1 adaptive standardization (sliding window W = 24 hours) and feature extraction, and the terrain roughness of the offshore wind farm area is calculated to be R ≈ 0.1, and the terrain roughness of the photovoltaic power plant area is R ≈ 0.05.

[0095] The improved U-Net model is expanded, the multi-branch feature extraction unit contains 64 channel convolutional layers, the spatial and temporal fusion module has 6 historical time points n = 6, and the attention enhancement module uses a 3x3 convolution kernel; In the physical constraint loss function, the reference wind speed y wind,ref The 95% quantile value of the wind speed of the training set (the selected offshore wind farm area is 10.2m / s), the theoretical maximum irradiance y rad,max is calculated by the solar elevation angle (the selected photovoltaic power plant area has a maximum of about 1000W / m 2 at noon). The training setting uses the Adam optimizer with an initial learning rate of 0.001, and iterates to the verification set R 2 is stable (about 30 rounds), and the batch size is set to 32 to balance the training efficiency and model performance. In order to highlight the technical advantages of the application, two types of existing technologies are selected for comparison: one is the traditional statistical downscaling method as Comparative Example 1, which uses a stepwise regression model combined with the statistical relationship between terrain factors and meteorological elements; The other is a conventional deep learning method as Comparative Example 2, which is a U-Net model without embedding physical constraints, and its loss function only uses MAE.

[0096] Through the above experimental design and process, the specific results of the offshore wind farm wind speed downscaling and the photovoltaic power plant light intensity downscaling are obtained, as shown in Tables 1 and 2 below and analyzed.

[0097] Table 1 Offshore wind farm wind speed downscaling results

[0098] Evaluation index Example 2 Comparative Example 1 Comparative Example 2 MAE (m / s) 0.32 0.87 0.45 RMSE (m / s) 0.48 1.12 0.67 [R 2 ]] 0.92 0.61 0.83 MAPE (%) 5.3 14.8 7.6

[0099] According to Table 1, in the wind speed downscaling task, the evaluation indexes of the application are significantly better than those of the traditional statistical method of Comparative Example 1 and the conventional U-Net model of Comparative Example 2. The MAE and RMSE of the application are the smallest, the R 2 The highest, and the MAPE is the lowest, indicating that the downscaling accuracy of the wind speed field is the highest, especially in capturing complex wind speed gradient changes such as sea-land wind conversion, which performs better, fully embodying the technical advantages of multi-branch feature extraction, spatio-temporal fusion and physical constraint mechanism in wind speed downscaling.

[0100] Table 2: Light intensity downscaling results of photovoltaic power plants

[0101] Evaluation index Example 2 Comparative Example 1 Comparative Example 2 MAE (W / m 2 )]]> 28.6 76.3 41.2 RMSE (W / m 2 ) 42.5 98.7 59.8 [R 2 ]] 0.91 0.58 0.82 MAPE (%) 6.2 16.5 8.9

[0102] According to Table 2, in the light intensity downscaling task, the evaluation indexes of the application are significantly better than those of the traditional statistical method of Comparative Example 1 and the conventional U-Net model of Comparative Example 2. The MAE and RMSE of the application are the smallest, the R 2 The highest, and the MAPE is the lowest, indicating that the downscaling accuracy of the light field is higher, especially in complex scenes such as cloudy disturbance, which effectively avoids extreme errors beyond the physical limit by virtue of the physical constraint mechanism.

[0103] Further, in the wind power plant and photovoltaic power plant scenes of the plain-near sea typical topography, the MAE and RMSE of the downscaling results of the application are significantly lower than those of the traditional method and the conventional deep learning method, the R 2 are not less than 0.91, indicating higher accuracy in restoring wind speed fine structure and light gradient; at the same time, the physical constraint mechanism effectively avoids extreme errors that violate meteorological laws, with MAPE controlled within 7%, having practical value for power generation prediction and operation decision-making in engineering applications, verifying the universality of the application in different resource types and terrain conditions.

[0104] Therefore, by using the above-mentioned meteorological downscaling method based on spatio-temporal fusion and physical constraint, through multi-branch feature extraction, spatio-temporal fusion, attention enhancement and loss function optimization by fusing terrain and physical rules, accurate downscaling of low-resolution meteorological data is realized, the prediction accuracy and physical rationality of high-resolution meteorological data in complex terrain areas and special meteorological scenes are improved, and physical attribution results can be output to clarify the influencing factors.

[0105] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above examples are only used to help understand the method of the present application and its core idea; meanwhile, for the general technical personnel in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In conclusion, the content of the present specification should not be understood as the limitation of the present application.

Claims

1. A weather downscaling method based on spatio-temporal fusion and physical constraints, characterized in that, The method comprises the following steps: S1, preprocessing multi-source meteorological related data and high-resolution meteorological true values to construct a training data set; the preprocessing comprises standardization and feature extraction; S2, constructing an improved U-Net model, the model obtains space-time features and multi-source auxiliary features through a multi-branch feature extraction unit, restores high-resolution information through fusion and decoding, and embeds an attention strengthening module to highlight key regional features; S3, training the model through the training data set, optimizing parameters by using a loss function that fuses terrain features and physical rules, and realizing differentiated constraint on prediction errors of complex regions and violations of physical laws; S4, inputting the target low-resolution data after preprocessing of step S1 into the trained model to output high-resolution meteorological data and physical attribution results.

2. The weather downscaling method based on spatio-temporal fusion and physical constraints according to claim 1, characterized in that, In step S1, the multi-source meteorological related data comprises low-resolution meteorological data, terrain data and multi-source observation data.

3. The weather downscaling method based on spatio-temporal fusion and physical constraints according to claim 2, characterized in that, The low-resolution meteorological data comprises one or more of wind speed, temperature and irradiance, the terrain data is digital elevation model data, and the multi-source observation data comprises satellite cloud images and radar echoes.

4. The weather downscaling method based on space-time fusion and physical constraints according to claim 1, characterized in that, In step S1, the standardization is adaptive normalization, specifically: where x' is the normalized data, x is the original data, μ(x W ) is the mean value in the sliding time window W, W is 24 hours, σ(x W ) is the standard deviation in the sliding time window W. The feature extraction comprises calculating terrain roughness, and the formula is: where R is the terrain roughness, β is the slope, and a is the aspect, is the gradient operator.

5. The method according to claim 1, wherein, In step S2, the improved U-Net model comprises a multi-branch feature extraction unit, a space-time fusion module, an attention strengthening module and a scene adaptive decoding unit; wherein the multi-branch feature extraction unit comprises three parallel branches: A meteorological space-time branch that extracts spatial features through 3×3 convolution and captures time sequence dynamics by combining a Transformer self-attention module, the time sequence attention weight calculation formula being: wherein a t is the timing attention weight, Softmax is the normalization function, Q t is the timing query vector, K t is the timing key vector, T is the matrix transpose symbol, d k is the vector dimension; A terrain branch that uses a residual block to process terrain derived features and enhance the feature extraction capability of complex terrain regions; A multi-source observation branch that extracts features from satellite cloud images through 7×7 convolution and radar echoes through 5×5 convolution to adapt to the distribution characteristics of different observation data.

6. The weather downscaling method based on spatio-temporal fusion and physical constraints according to claim 5, characterized in that, The space-time fusion module is used to fuse the features output by the multi-branch feature extraction unit, and realizes associated learning of spatial and time sequence features through a cross-attention matrix, the formula being: wherein M st is a spatio-temporal cross-attention weight matrix, F s is a spatial feature vector, T is a matrix transpose symbol, F t is a current time sequence feature vector, σ st is a spatio-temporal feature covariance, which is obtained by self-adaptive learning from training data; n is a historical time quantity, and is 6; F t,i is an i-th historical time sequence feature vector.

7. The method according to claim 5, wherein, The attention strengthening module is used to highlight key regional features, and adjusts the feature importance of different regions through spatial attention weights, the formula being: where W s is the spatial attention weight, Sigmoid is the activation function, Conv is the 3 × 3 convolution operation, Cat is the feature concatenation operation, MaxPool is the max-pooling operation, AvgPool is the average-pooling operation, F is the input feature map, is the terrain roughness gradient.

8. The weather downscaling method based on spatio-temporal fusion and physical constraints according to claim 5, characterized in that, The scene adaptive decoding unit is used to dynamically adjust the decoding strategy according to the meteorological scene, determines the branch activation function through scene classification, and the formula being: where P c is the activation probability of the cth scenario, exp is the exponential function, MLP is the multi-layer perception mechanism, F stat is the statistical feature vector of the current data, F stat,i is the reference statistical feature of the ith scenario.

9. The weather downscaling method based on spatio-temporal fusion and physical constraints according to claim 1, characterized in that, In step S3, the loss function that fuses terrain features and physical rules has the formula: where L is the comprehensive loss value, y is the high-resolution meteorological true value, is the predicted value, R is the terrain roughness, tanh is the hyperbolic tangent function, y wind is the predicted wind speed, y wind,ref is the reference wind speed, y rad is the predicted irradiance, y rad,max is the theoretical maximum irradiance, which is calculated from the solar elevation angle.

10. The method according to claim 1, wherein, In step S4, the terrain contribution degree of the physical attribution result has the formula: where C R is the terrain contribution, with a value range of 0-1; W s is the spatial attention weight, is the terrain roughness gradient, y is the high-resolution meteorological true value, is the predicted value, is the absolute error between the predicted value and the true value.

Images